Rice Pest Recognition Method Based on Improved YOLOv8

Rapid and accurate identification of rice pests is a prerequisite for the control of rice pests. Due to the problem of miss-detection and misdetection of small targets in rice pests, this paper constructs the rice pest dataset RiPest and proposes an improved YOLOv8 algorithm Gi-YOLOv8. Firstly, the GAM attention mechanism is added to the YOLOv8 backbone network to preserve the correlation between spatial and channel information to improve the attention to small targets, and secondly, the BiFPN network is fused in the Neck part to achieve weighted fusion and bi-directional cross-scalar connectivity, which improves the model detection capability. Experiments show that the Gi- YOLOv8 model has an accuracy of 92.5% in the self-made RiPest rice pest dataset in this paper, which is a 1.3% improvement compared to before the improvement, and the recall and mAP50 are improved by 0.6% and 0.2%, respectively. The experimental results show that the Gi-YOLOv8 model can enhance the detection of small-target rice pests and provide an effective decision support tool for rice pest control.